A Framework for Designing Efficient Deep Learning-Based Genomic Basecallers

Abstract

Nanopore sequencing is a widely-used high-throughput genome sequencing technology that can sequence long fragments of a genome. Nanopore sequencing generates noisy electrical signals that need to be converted into a standard string of DNA nucleotide bases (i.e., A, C, G, T) using a computational step called basecalling. The accuracy and speed of basecalling have critical implications for every subsequent step in genome analysis. Currently, basecallers are mainly based on deep learning techniques to provide high sequencing accuracy without considering the compute demands of such tools. We observe that state-of-the-art basecallers (i.e., Guppy, Bonito) are slow, inefficient, and memory-hungry as researchers have adapted deep learning models from other domains without specialization to the basecalling purpose. Our goal is to make basecalling highly efficient and fast by building the first framework for specializing and optimizing machine learning-based basecaller. We introduce RUBICON, a framework to develop hardware-optimized basecallers. RUBICON consists of two novel machine-learning techniques that are specifically designed for basecalling. First, we introduce the quantization-aware basecalling neural architecture search (QABAS) framework to specialize the basecalling neural network architecture for a given hardware acceleration platform while jointly exploring and finding the best bit-width precision for each neural network layer. Second, we develop SkipClip, the first technique to remove all the skip connections present in modern basecallers to greatly reduce resource and storage requirements without any loss in basecalling accuracy. We demonstrate the benefits of QABAS and SkipClip by developing RUBICALL, the first hardware-optimized basecaller that performs fast and accurate basecalling. Our experimental results on state-of-the-art computing systems show that RUBICALL is a fast, accurate and hardware-friendly, mixed-precision basecaller. Compared to a highly-accurate state-of-the-art basecaller, RUBICALL provides a 16.56x speedup without losing accuracy, while also achieving a 6.88x and 2.94x reduction in neural network model size and the number of parameters, respectively. Compared to the fastest state-of-the-art basecaller, RUBICALL provides a 3.19x speedup with 2.97% higher accuracy. We show that QABAS and SkipClip can help researchers develop hardware-optimized basecallers that are superior to expert-designed models.

RawHash: Enabling Fast and Accurate Real-Time Analysis of Raw Nanopore Signals for Large Genomes

“…One of the immediate steps after generating raw nanopore signals is their translation to their corresponding DNA bases as sequences of characters with a computationallyintensive step, basecalling. Basecalling approaches are usually computationally costly and consume significant energy as they use complex deep learning models [26][27][28][29][30][31][32][33][34][35][36][37][38]. Although we do not evaluate in this work, we expect that RawHash can be used as a low-cost filter to eliminate the reads that are unlikely to be useful in downstream analysis, which can reduce the overall workload of basecallers and further downstream analysis.…”

TargetCall: Eliminating the Wasted Computation in Basecalling via Pre-Basecalling Filtering

“…Modern deep learning-based basecallers [20, 21, 36, 38, 39, 48, 52–54] incorporate skip connections to help mitigate the vanishing gradient and saturation problems [55]. Removing skip connections has a higher impact on basecalling accuracy.…”

“…However, these deep learning models use millions of model parameters that makes basecalling computationally extensive. Recent works propose algorithmic optimizations [33, 48] and hardware accelerators [85] to improve the performance of basecallers. These works accelerate the basecalling step without eliminating the wasted computation in basecalling.…”

“…We develop LightCall by modifying the state-of-the-art basecaller Bonito's architecture in three ways: 1) reducing the channel sizes of convolution layers, 2) removing the skip connections, and 3) reducing the number of basic computation blocks. Prior work [48] shows that Bonito's model is over provisioned, and we can maintain very high accuracy with reduced model sizes. We leverage this finding in TargetCall to reduce the channel sizes of convolution layers and number of basic computation blocks of LightCall.…”

“…Our LightCall architecture is composed of 18 convolution blocks containing ∼292 thousand model parameters (∼33.35× lower parameters than Bonito). Modern deep learning-based basecallers [20,21,36,38,39,48,[52][53][54] incorporate skip connections to help mitigate the vanishing gradient and saturation problems [55]. Removing skip connections has a higher impact on basecalling accuracy.…”